Age bias in survey sampling and implications for estimating HIV prevalence in men who have sex with men: insights from mathematical modelling

Abstract

Respondent-driven sampling (RDS) is widely used to estimate HIV prevalence in men who have sex with men (MSM). Mathematical models that are calibrated to these data may be compromised if they fail to account for selection biases in RDS surveys. To quantify the potential extent of this bias, an agent-based model of HIV in South Africa was calibrated to HIV prevalence and sexual behaviour data from South African studies of MSM, first reweighting the modelled MSM population to match the younger age profile of the RDS surveys (age-adjusted analysis) and then without reweighting (unadjusted analysis). The model estimated a median HIV prevalence in South African MSM in 2015 of 34.6% (inter-quartile range (IQR): 31.4-37.2%) in the age-adjusted analysis, compared with 26.1% (IQR: 24.1-28.4%) in the unadjusted analysis. The median lifetime risk of acquiring HIV in exclusively homosexual men was 88% (IQR: 82-92%) in the age-adjusted analysis, compared with 76% (IQR: 64-85%) in the unadjusted analysis. These results suggest that RDS studies may under-estimate the exceptionally high HIV prevalence rates in South African MSM because of over-sampling of younger MSM. Mathematical models that are calibrated to these data need to control for likely over-sampling of younger MSM.

MicroCOSM: a model of social and structural drivers of HIV and interventions to reduce HIV incidence in high-risk populations in South Africa

“…Of the men who are gay or bisexual, 77% are assumed to be in the bisexual category. This estimate is the average of the 100 best-fitting parameters obtained when the model was previously fitted to South African sexual behaviour data from MSM [160]. The data are from respondent-driven sampling (RDS) studies, in which the proportion of MSM who reported having ever had sex with women varied between 36% and 87% (average value of 58%) [131, 135, 161–163].…”

“…To reflect this heterogeneity, we sample the a i values from a beta distribution with mean 0.23 and standard deviation 0.09 and sample the b i values from a normal distribution with mean - 0.015 and standard deviation 0.16. As before, each parameter estimate is calculated as the average of the 100 best-fitting parameters obtained when the model was previously fitted to South African sexual behaviour data from MSM [160]. The model thus suggests that men in the bisexual category tend to have more female partners than male partners, and that there is a slight tendency towards relatively more female partners as men get older (although with substantial variation between individuals).…”

“…We therefore reduce the heterosexual marriage rates by a constant factor when modelling the rate at which short-term relationships between MSM become marital. This factor has been set to 0.48, the average of the 100 best-fitting parameters obtained when the model was fitted to behavioural data from South African surveys of MSM [160].…”

“…This is almost certainly an under-estimate of the duration that would be expected in short-term relationships, (a) because men could not report durations of more than 6 months, and (b) because reporting included casual sex encounters (which would in most cases have been of very short duration). The average duration of short-term relationships in MSM has been set to 0.48 years (5.8 months), the average of the 100 best-fitting parameters obtained when the model was fitted to behavioural data from surveys of South African MSM [160].…”

“…The values assigned to the remaining parameters are summarized in Table 4.7.1. All of these parameters have been calculated as the average of the 100 best-fitting parameters obtained when the model was fitted to behavioural and HIV prevalence data from South African surveys of MSM [160]. As might be expected, low-risk MSM are estimated to have a lower rate of entry into casual sex than high-risk MSM.…”